An optimized workflow scheduling algorithm on CPU-GPU heterogeneous systems

Abstract

Abstract Aiming at the workflow scheduling problem on CPU-GPU heterogeneous systems, this article proposes a workflow scheduling algorithm that optimizes task priority and processor selection phase. This article focuses on the relationship between processor processing time and task acceleration ratio, and uses optimistic finish time table. This article estimates each task’s communication cost and computation cost on the basis of each task’s acceleration ratio ri. Then the algorithm calculates the task’s priority ranku based on communication cost and computation cost. In processor selection phase, this article takes the Earliest Finish Time (EFT) difference between the processor with the fastest task and the processor selected based on Heterogeneous Earliest Finish Time (HEFT) algorithm as the judgment condition k. If the difference between the earliest completion time of all child tasks is greater than k, the algorithm selects the processor selected by the HEFT algorithm; otherwise, the algorithm selects another one. This article uses the computational shared facility’s kernel timer for simulation experiment, which bases on the University of Manchester’s high performance computing cluster. Based on the simulation results, the proposed algorithm can reduce the maximum completion time and energy consumption, and improve the acceleration ratio. Compared with the HEFT algorithm, the task maximum completion time and energy consumption are reduced by 10% and 5% respectively, and the acceleration ratio is improved by 3%.